Thursday, April 17, 2008

More Image Editing

Yin Li , Jian Sun , Chi-Keung Tang , Heung-Yeung Shum. 2004. Lazy snapping. Proc. ACM SIGGRAPH, 303-308.

Elder, J.H. Goldberg, R.M. 2001. Image editing in the contour domain. IEEE Trans. Pattern Analysis and Machine Intelligence, 23, 3, 291-296.


Unknown said...

Elder's paper:
I didn't find this paper terribly enlightening. It seems to be more of a "here's how to combine other work into a product." Thus, I didn't see a whole lot of details past the high-level overview.

Lazy Snapping:
This is a fairly straightforward application of graphcuts from what I can see. It's another one of those "why did it take us until 2004" to figure this out. I think the best idea here is probably the simplest: the multi-level approach to identifying the object.

Stuart Heinrich said...

Lazy snapping - magnetic lasso in photoshop is pretty much useless, so the "state of the art" (in the industry, not necessarily in research literature) could be better described as drawing the outline by hand (imo). It looks like the Lazy snapping does a pretty good job. I like the interactive design. Formulating the graph cut minimization on top of watershed segmentation was a good idea to improve performance. I just wish that their results had not been so jpeg compressed. When zooming into the cutout boundary, it is not possible to tell if the edges are anti aliased or not. It seems to me from the description of their method that it would not produce anti-aliased edges. If this is the case then it would probably still be inferior to manual cutout in terms of quality, because a manual segmentation does provide anti-aliased edges and that is extremely important when compositing a foreground object into a scene that has a different background color.

Stuart Heinrich said...

Image Editing in Contour Domain

Interesting to see this paper as it illustrates another use of the Poisson equation we talked about last time.

The concept of being able to delete edges and reconstruct the image is neat. Unfortunately their map does not preserve enough fine detail, so that all of the reconstructions suffer an undesirable loss of quality.

I think that a similar goal could be achieved by simply masking off regions in the IMAGE domain and saying that you wanted to get rid of edges there...but this would not have the problem of losing detail everywhere. There's no reason to go into the "contour domain".